Senone tree representation and evaluation

摘要: A speech recognition method provides improved modeling in accuracy using hidden Markov models. During training, the creates a senone tree for each state of phoneme encountered data set training words. All output distributions received selected words are clustered together root node tree. Each beginning with is divided into two nodes by asking linguistic questions regarding phonemes immediately to left and right central triphone. At predetermined point, creation stops, resulting leaves representing known as senones. The trees allow all possible triphones be mapped sequence senones simply traversing associated As result, unseen not can modeled created actually found data.